IBM SPSS Categories 24 IBM

Transcription

1 IBM SPSS Categories 24 IBM

2 Note Before using this information and the product it supports, read the information in Notices on page 55. Product Information This edition applies to ersion 24, release 0, modification 0 of IBM SPSS Statistics and to all subsequent releases and modifications until otherwise indicated in new editions.

3 Contents Chapter 1. Introduction to Optimal Scaling Procedures for Categorical Data. 1 What Is Optimal Scaling? Why Use Optimal Scaling? Optimal Scaling Leel and Measurement Leel... 2 Selecting the Optimal Scaling Leel Transformation Plots Category Codes Which Procedure Is Best for Your Application?... 5 Categorical Regression Categorical Principal Components Analysis... 6 Nonlinear Canonical Correlation Analysis... 6 Correspondence Analysis Multiple Correspondence Analysis Multidimensional Scaling Multidimensional Unfolding Aspect Ratio in Optimal Scaling Charts Chapter 2. Categorical Regression (CATREG) Define Scale in Categorical Regression Categorical Regression Discretization Categorical Regression Missing Values Categorical Regression Options Categorical Regression Regularization Categorical Regression Output Categorical Regression Sae Categorical Regression Transformation Plots CATREG Command Additional Features Chapter 3. Categorical Principal Components Analysis (CATPCA) Define Scale and Weight in CATPCA Categorical Principal Components Analysis Discretization Categorical Principal Components Analysis Missing Values Categorical Principal Components Analysis Options 19 Categorical Principal Components Analysis Output 21 Categorical Principal Components Analysis Sae.. 21 Categorical Principal Components Analysis Object Plots Categorical Principal Components Analysis Category Plots Categorical Principal Components Analysis Loading Plots Categorical Principal Components Analysis Bootstrap CATPCA Command Additional Features Chapter 4. Nonlinear Canonical Correlation Analysis (OVERALS) Define Range and Scale Define Range Nonlinear Canonical Correlation Analysis Options 27 OVERALS Command Additional Features Chapter 5. Correspondence Analysis 29 Define Row Range in Correspondence Analysis.. 30 Define Column Range in Correspondence Analysis 30 Correspondence Analysis Model Correspondence Analysis Statistics Correspondence Analysis Plots CORRESPONDENCE Command Additional Features Chapter 6. Multiple Correspondence Analysis Define Variable Weight in Multiple Correspondence Analysis Multiple Correspondence Analysis Discretization.. 36 Multiple Correspondence Analysis Missing Values 36 Multiple Correspondence Analysis Options Multiple Correspondence Analysis Output Multiple Correspondence Analysis Sae Multiple Correspondence Analysis Object Plots.. 38 Multiple Correspondence Analysis Variable Plots.. 39 MULTIPLE CORRESPONDENCE Command Additional Features Chapter 7. Multidimensional Scaling (PROXSCAL) Proximities in Matrices across Columns Proximities in Columns Proximities in One Column Create Proximities from Data Create Measure from Data Define a Multidimensional Scaling Model Multidimensional Scaling Restrictions Multidimensional Scaling Options Multidimensional Scaling Plots, Version Multidimensional Scaling Plots, Version Multidimensional Scaling Output PROXSCAL Command Additional Features Chapter 8. Multidimensional Unfolding (PREFSCAL) Define a Multidimensional Unfolding Model Multidimensional Unfolding Restrictions Multidimensional Unfolding Options Multidimensional Unfolding Plots Multidimensional Unfolding Output PREFSCAL Command Additional Features Notices Trademarks Index iii

4 i IBM SPSS Categories 24

5 Chapter 1. Introduction to Optimal Scaling Procedures for Categorical Data Categories procedures use optimal scaling to analyze data that are difficult or impossible for standard statistical procedures to analyze. This chapter describes what each procedure does, the situations in which each procedure is most appropriate, the relationships between the procedures, and the relationships of these procedures to their standard statistical counterparts. Note: These procedures and their implementation in IBM SPSS Statistics were deeloped by the Data Theory Scaling System Group (DTSS), consisting of members of the departments of Education and Psychology, Faculty of Social and Behaioral Sciences, Leiden Uniersity. What Is Optimal Scaling? The idea behind optimal scaling is to assign numerical quantifications to the categories of each ariable, thus allowing standard procedures to be used to obtain a solution on the quantified ariables. The optimal scale alues are assigned to categories of each ariable based on the optimizing criterion of the procedure in use. Unlike the original labels of the nominal or ordinal ariables in the analysis, these scale alues hae metric properties. In most Categories procedures, the optimal quantification for each scaled ariable is obtained through an iteratie method called alternating least squares in which, after the current quantifications are used to find a solution, the quantifications are updated using that solution. The updated quantifications are then used to find a new solution, which is used to update the quantifications, and so on, until some criterion is reached that signals the process to stop. Why Use Optimal Scaling? Categorical data are often found in marketing research, surey research, and research in the social and behaioral sciences. In fact, many researchers deal almost exclusiely with categorical data. While adaptations of most standard models exist specifically to analyze categorical data, they often do not perform well for datasets that feature: Too few obserations Too many ariables Too many alues per ariable By quantifying categories, optimal scaling techniques aoid problems in these situations. Moreoer, they are useful een when specialized techniques are appropriate. Rather than interpreting parameter estimates, the interpretation of optimal scaling output is often based on graphical displays. Optimal scaling techniques offer excellent exploratory analyses, which complement other IBM SPSS Statistics models well. By narrowing the focus of your inestigation, isualizing your data through optimal scaling can form the basis of an analysis that centers on interpretation of model parameters. Copyright IBM Corporation 1989,

6 Optimal Scaling Leel and Measurement Leel This can be a ery confusing concept when you first use Categories procedures. When specifying the leel, you specify not the leel at which ariables are measured but the leel at which they are scaled. The idea is that the ariables to be quantified may hae nonlinear relations regardless of how they are measured. For Categories purposes, there are three basic leels of measurement: The nominal leel implies that a ariable's alues represent unordered categories. Examples of ariables that might be nominal are region, zip code area, religious affiliation, and multiple choice categories. The ordinal leel implies that a ariable's alues represent ordered categories. Examples include attitude scales representing degree of satisfaction or confidence and preference rating scores. The numerical leel implies that a ariable's alues represent ordered categories with a meaningful metric so that distance comparisons between categories are appropriate. Examples include age in years and income in thousands of dollars. For example, suppose that the ariables region, job, and age are coded as shown in the following table. Table 1. Coding scheme for region, job, and age Region code Region alue Job code Job alue Age 1 North 1 intern 20 2 South 2 sales rep 22 3 East 3 manager 25 4 West 27 The alues shown represent the categories of each ariable. Region would be a nominal ariable. There are four categories of region, with no intrinsic ordering. Values 1 through 4 simply represent the four categories; the coding scheme is completely arbitrary. Job, on the other hand, could be assumed to be an ordinal ariable. The original categories form a progression from intern to manager. Larger codes represent a job higher on the corporate ladder. Howeer, only the order information is known--nothing can be said about the distance between adjacent categories. In contrast, age could be assumed to be a numerical ariable. In the case of age, the distances between the alues are intrinsically meaningful. The distance between 20 and 22 is the same as the distance between 25 and 27, while the distance between 22 and 25 is greater than either of these. Selecting the Optimal Scaling Leel It is important to understand that there are no intrinsic properties of a ariable that automatically predefine what optimal scaling leel you should specify for it. You can explore your data in any way that makes sense and makes interpretation easier. By analyzing a numerical-leel ariable at the ordinal leel, for example, the use of a nonlinear transformation may allow a solution in fewer dimensions. The following two examples illustrate how the "obious" leel of measurement might not be the best optimal scaling leel. Suppose that a ariable sorts objects into age groups. Although age can be scaled as a numerical ariable, it may be true that for people younger than 25 safety has a positie relation with age, whereas for people older than 60 safety has a negatie relation with age. In this case, it might be better to treat age as a nominal ariable. As another example, a ariable that sorts persons by political preference appears to be essentially nominal. Howeer, if you order the parties from political left to political right, you might want the quantification of parties to respect this order by using an ordinal leel of analysis. 2 IBM SPSS Categories 24

7 Een though there are no predefined properties of a ariable that make it exclusiely one leel or another, there are some general guidelines to help the noice user. With single-nominal quantification, you don't usually know the order of the categories but you want the analysis to impose one. If the order of the categories is known, you should try ordinal quantification. If the categories are unorderable, you might try multiple-nominal quantification. Transformation Plots The different leels at which each ariable can be scaled impose different restrictions on the quantifications. Transformation plots illustrate the relationship between the quantifications and the original categories resulting from the selected optimal scaling leel. For example, a linear transformation plot results when a ariable is treated as numerical. Variables treated as ordinal result in a nondecreasing transformation plot. Transformation plots for ariables treated nominally that are U-shaped (or the inerse) display a quadratic relationship. Nominal ariables could also yield transformation plots without apparent trends by changing the order of the categories completely. The following figure displays a sample transformation plot. Transformation plots are particularly suited to determining how well the selected optimal scaling leel performs. If seeral categories receie similar quantifications, collapsing these categories into one category may be warranted. Alternatiely, if a ariable treated as nominal receies quantifications that display an increasing trend, an ordinal transformation may result in a similar fit. If that trend is linear, numerical treatment may be appropriate. Howeer, if collapsing categories or changing scaling leels is warranted, the analysis will not change significantly. Category Codes Some care should be taken when coding categorical ariables because some coding schemes may yield unwanted output or incomplete analyses. Possible coding schemes for job are displayed in the following table. Table 2. Alternatie coding schemes for job Category A B C D intern sales rep manager Some Categories procedures require that the range of eery ariable used be defined. Any alue outside this range is treated as a missing alue. The minimum category alue is always 1. The maximum category alue is supplied by the user. This alue is not the number of categories for a ariable it is the largest category alue. For example, in the table, scheme A has a maximum category alue of 3 and scheme B has a maximum category alue of 7, yet both schemes code the same three categories. The ariable range determines which categories will be omitted from the analysis. Any categories with codes outside the defined range are omitted from the analysis. This is a simple method for omitting categories but can result in unwanted analyses. An incorrectly defined maximum category can omit alid categories from the analysis. For example, for scheme B, defining the maximum category alue to be 3 indicates that job has categories coded from 1 to 3; the manager category is treated as missing. Because no category has actually been coded 3, the third category in the analysis contains no cases. If you wanted to omit all manager categories, this analysis would be appropriate. Howeer, if managers are to be included, the maximum category must be defined as 7, and missing alues must be coded with alues aboe 7 or below 1. For ariables treated as nominal or ordinal, the range of the categories does not affect the results. For nominal ariables, only the label and not the alue associated with that label is important. For ordinal ariables, the order of the categories is presered in the quantifications; the category alues themseles Chapter 1. Introduction to Optimal Scaling Procedures for Categorical Data 3

8 are not important. All coding schemes resulting in the same category ordering will hae identical results. For example, the first three schemes in the table are functionally equialent if job is analyzed at an ordinal leel. The order of the categories is identical in these schemes. Scheme D, on the other hand, inerts the second and third categories and will yield different results than the other schemes. Although many coding schemes for a ariable are functionally equialent, schemes with small differences between codes are preferred because the codes hae an impact on the amount of output produced by a procedure. All categories coded with alues between 1 and the user-defined maximum are alid. If any of these categories are empty, the corresponding quantifications will be either system-missing or 0, depending on the procedure. Although neither of these assignments affect the analyses, output is produced for these categories. Thus, for scheme B, job has four categories that receie system-missing alues. For scheme C, there are also four categories receiing system-missing indicators. In contrast, for scheme A there are no system-missing quantifications. Using consecutie integers as codes for ariables treated as nominal or ordinal results in much less output without affecting the results. Coding schemes for ariables treated as numerical are more restricted than the ordinal case. For these ariables, the differences between consecutie categories are important. The following table displays three coding schemes for age. Table 3. Alternatie coding schemes for age Category A B C Any recoding of numerical ariables must presere the differences between the categories. Using the original alues is one method for ensuring preseration of differences. Howeer, this can result in many categories haing system-missing indicators. For example, scheme A employs the original obsered alues. For all Categories procedures except for Correspondence Analysis, the maximum category alue is 27 and the minimum category alue is set to 1. The first 19 categories are empty and receie system-missing indicators. The output can quickly become rather cumbersome if the maximum category is much greater than 1 and there are many empty categories between 1 and the maximum. To reduce the amount of output, recoding can be done. Howeer, in the numerical case, the Automatic Recode facility should not be used. Coding to consecutie integers results in differences of 1 between all consecutie categories, and, as a result, all quantifications will be equally spaced. The metric characteristics deemed important when treating a ariable as numerical are destroyed by recoding to consecutie integers. For example, scheme C in the table corresponds to automatically recoding age. The difference between categories 22 and 25 has changed from three to one, and the quantifications will reflect the latter difference. An alternatie recoding scheme that preseres the differences between categories is to subtract the smallest category alue from eery category and add 1 to each difference. Scheme B results from this transformation. The smallest category alue, 20, has been subtracted from each category, and 1 was added to each result. The transformed codes hae a minimum of 1, and all differences are identical to the original data. The maximum category alue is now 8, and the zero quantifications before the first nonzero quantification are all eliminated. Yet, the nonzero quantifications corresponding to each category resulting from scheme B are identical to the quantifications from scheme A. 4 IBM SPSS Categories 24

9 Which Procedure Is Best for Your Application? The techniques embodied in four of these procedures (Correspondence Analysis, Multiple Correspondence Analysis, Categorical Principal Components Analysis, and Nonlinear Canonical Correlation Analysis) fall into the general area of multiariate data analysis known as dimension reduction. That is, relationships between ariables are represented in a few dimensions say two or three as often as possible. This enables you to describe structures or patterns in the relationships that would be too difficult to fathom in their original richness and complexity. In market research applications, these techniques can be a form of perceptual mapping. A major adantage of these procedures is that they accommodate data with different leels of optimal scaling. Categorical Regression describes the relationship between a categorical response ariable and a combination of categorical predictor ariables. The influence of each predictor ariable on the response ariable is described by the corresponding regression weight. As in the other procedures, data can be analyzed with different leels of optimal scaling. Multidimensional Scaling and Multidimensional Unfolding describe relationships between objects in a low-dimensional space, using the proximities between the objects. Following are brief guidelines for each of the procedures: Use Categorical Regression to predict the alues of a categorical dependent ariable from a combination of categorical independent ariables. Use Categorical Principal Components Analysis to account for patterns of ariation in a single set of ariables of mixed optimal scaling leels. Use Nonlinear Canonical Correlation Analysis to assess the extent to which two or more sets of ariables of mixed optimal scaling leels are correlated. Use Correspondence Analysis to analyze two-way contingency tables or data that can be expressed as a two-way table, such as brand preference or sociometric choice data. Use Multiple Correspondence Analysis to analyze a categorical multiariate data matrix when you are willing to make no stronger assumption that all ariables are analyzed at the nominal leel. Use Multidimensional Scaling to analyze proximity data to find a least-squares representation of a single set of objects in a low-dimensional space. Use Multidimensional Unfolding to analyze proximity data to find a least-squares representation of two sets of objects in a low-dimensional space. Categorical Regression The use of Categorical Regression is most appropriate when the goal of your analysis is to predict a dependent (response) ariable from a set of independent (predictor) ariables. As with all optimal scaling procedures, scale alues are assigned to each category of eery ariable such that these alues are optimal with respect to the regression. The solution of a categorical regression maximizes the squared correlation between the transformed response and the weighted combination of transformed predictors. Relation to other Categories procedures. Categorical regression with optimal scaling is comparable to optimal scaling canonical correlation analysis with two sets, one of which contains only the dependent ariable. In the latter technique, similarity of sets is deried by comparing each set to an unknown ariable that lies somewhere between all of the sets. In categorical regression, similarity of the transformed response and the linear combination of transformed predictors is assessed directly. Relation to standard techniques. In standard linear regression, categorical ariables can either be recoded as indicator ariables or be treated in the same fashion as interal leel ariables. In the first approach, the model contains a separate intercept and slope for each combination of the leels of the categorical ariables. This results in a large number of parameters to interpret. In the second approach, only one parameter is estimated for each ariable. Howeer, the arbitrary nature of the category codings makes generalizations impossible. Chapter 1. Introduction to Optimal Scaling Procedures for Categorical Data 5

10 If some of the ariables are not continuous, alternatie analyses are aailable. If the response is continuous and the predictors are categorical, analysis of ariance is often employed. If the response is categorical and the predictors are continuous, logistic regression or discriminant analysis may be appropriate. If the response and the predictors are both categorical, loglinear models are often used. Regression with optimal scaling offers three scaling leels for each ariable. Combinations of these leels can account for a wide range of nonlinear relationships for which any single "standard" method is ill-suited. Consequently, optimal scaling offers greater flexibility than the standard approaches with minimal added complexity. In addition, nonlinear transformations of the predictors usually reduce the dependencies among the predictors. If you compare the eigenalues of the correlation matrix for the predictors with the eigenalues of the correlation matrix for the optimally scaled predictors, the latter set will usually be less ariable than the former. In other words, in categorical regression, optimal scaling makes the larger eigenalues of the predictor correlation matrix smaller and the smaller eigenalues larger. Categorical Principal Components Analysis The use of Categorical Principal Components Analysis is most appropriate when you want to account for patterns of ariation in a single set of ariables of mixed optimal scaling leels. This technique attempts to reduce the dimensionality of a set of ariables while accounting for as much of the ariation as possible. Scale alues are assigned to each category of eery ariable so that these alues are optimal with respect to the principal components solution. Objects in the analysis receie component scores based on the quantified data. Plots of the component scores reeal patterns among the objects in the analysis and can reeal unusual objects in the data. The solution of a categorical principal components analysis maximizes the correlations of the object scores with each of the quantified ariables for the number of components (dimensions) specified. An important application of categorical principal components is to examine preference data, in which respondents rank or rate a number of items with respect to preference. In the usual IBM SPSS Statistics data configuration, rows are indiiduals, columns are measurements for the items, and the scores across rows are preference scores (on a 0 to 10 scale, for example), making the data row-conditional. For preference data, you may want to treat the indiiduals as ariables. Using the Transpose procedure, you can transpose the data. The raters become the ariables, and all ariables are declared ordinal. There is no objection to using more ariables than objects in CATPCA. Relation to other Categories procedures. If all ariables are declared multiple nominal, categorical principal components analysis produces an analysis equialent to a multiple correspondence analysis run on the same ariables. Thus, categorical principal components analysis can be seen as a type of multiple correspondence analysis in which some of the ariables are declared ordinal or numerical. Relation to standard techniques. If all ariables are scaled on the numerical leel, categorical principal components analysis is equialent to standard principal components analysis. More generally, categorical principal components analysis is an alternatie to computing the correlations between non-numerical scales and analyzing them using a standard principal components or factor-analysis approach. Naie use of the usual Pearson correlation coefficient as a measure of association for ordinal data can lead to nontriial bias in estimation of the correlations. Nonlinear Canonical Correlation Analysis Nonlinear Canonical Correlation Analysis is a ery general procedure with many different applications. The goal of nonlinear canonical correlation analysis is to analyze the relationships between two or more sets of ariables instead of between the ariables themseles, as in principal components analysis. For example, you may hae two sets of ariables, where one set of ariables might be demographic background items on a set of respondents and a second set might be responses to a set of attitude items. The scaling leels in the analysis can be any mix of nominal, ordinal, and numerical. Optimal scaling 6 IBM SPSS Categories 24

11 canonical correlation analysis determines the similarity among the sets by simultaneously comparing the canonical ariables from each set to a compromise set of scores assigned to the objects. Relation to other Categories procedures. If there are two or more sets of ariables with only one ariable per set, optimal scaling canonical correlation analysis is equialent to optimal scaling principal components analysis. If all ariables in a one-ariable-per-set analysis are multiple nominal, optimal scaling canonical correlation analysis is equialent to multiple correspondence analysis. If there are two sets of ariables, one of which contains only one ariable, optimal scaling canonical correlation analysis is equialent to categorical regression with optimal scaling. Relation to standard techniques. Standard canonical correlation analysis is a statistical technique that finds a linear combination of one set of ariables and a linear combination of a second set of ariables that are maximally correlated. Gien this set of linear combinations, canonical correlation analysis can find subsequent independent sets of linear combinations, referred to as canonical ariables, up to a maximum number equal to the number of ariables in the smaller set. If there are two sets of ariables in the analysis and all ariables are defined to be numerical, optimal scaling canonical correlation analysis is equialent to a standard canonical correlation analysis. Although IBM SPSS Statistics does not hae a canonical correlation analysis procedure, many of the releant statistics can be obtained from multiariate analysis of ariance. Optimal scaling canonical correlation analysis has arious other applications. If you hae two sets of ariables and one of the sets contains a nominal ariable declared as single nominal, optimal scaling canonical correlation analysis results can be interpreted in a similar fashion to regression analysis. If you consider the ariable to be multiple nominal, the optimal scaling analysis is an alternatie to discriminant analysis. Grouping the ariables in more than two sets proides a ariety of ways to analyze your data. Correspondence Analysis The goal of correspondence analysis is to make biplots for correspondence tables. In a correspondence table, the row and column ariables are assumed to represent unordered categories; therefore, the nominal optimal scaling leel is always used. Both ariables are inspected for their nominal information only. That is, the only consideration is the fact that some objects are in the same category while others are not. Nothing is assumed about the distance or order between categories of the same ariable. One specific use of correspondence analysis is the analysis of two-way contingency tables. If a table has r actie rows and c actie columns, the number of dimensions in the correspondence analysis solution is the minimum of r minus 1 or c minus 1, whicheer is less. In other words, you could perfectly represent the row categories or the column categories of a contingency table in a space of dimensions. Practically speaking, howeer, you would like to represent the row and column categories of a two-way table in a low-dimensional space, say two dimensions, for the reason that two-dimensional plots are more easily comprehensible than multidimensional spatial representations. When fewer than the maximum number of possible dimensions is used, the statistics produced in the analysis describe how well the row and column categories are represented in the low-dimensional representation. Proided that the quality of representation of the two-dimensional solution is good, you can examine plots of the row points and the column points to learn which categories of the row ariable are similar, which categories of the column ariable are similar, and which row and column categories are similar to each other. Relation to other Categories procedures. Simple correspondence analysis is limited to two-way tables. If there are more than two ariables of interest, you can combine ariables to create interaction ariables. For example, for the ariables region, job, and age, you can combine region and job to create a new ariable rejob with the 12 categories shown in the following table. This new ariable forms a two-way table with age (12 rows, 4 columns), which can be analyzed in correspondence analysis. Chapter 1. Introduction to Optimal Scaling Procedures for Categorical Data 7

12 Table 4. Combinations of region and job Category code Category definition Category code Category definition 1 North, intern 7 East, intern 2 North, sales rep 8 East, sales rep 3 North, manager 9 East, manager 4 South, intern 10 West, intern 5 South, sales rep 11 West, sales rep 6 South, manager 12 West, manager One shortcoming of this approach is that any pair of ariables can be combined. We can combine job and age, yielding another 12-category ariable. Or we can combine region and age, which results in a new 16-category ariable. Each of these interaction ariables forms a two-way table with the remaining ariable. Correspondence analyses of these three tables will not yield identical results, yet each is a alid approach. Furthermore, if there are four or more ariables, two-way tables comparing an interaction ariable with another interaction ariable can be constructed. The number of possible tables to analyze can get quite large, een for a few ariables. You can select one of these tables to analyze, or you can analyze all of them. Alternatiely, the Multiple Correspondence Analysis procedure can be used to examine all of the ariables simultaneously without the need to construct interaction ariables. Relation to standard techniques. The Crosstabs procedure can also be used to analyze contingency tables, with independence as a common focus in the analyses. Howeer, een in small tables, detecting the cause of departures from independence may be difficult. The utility of correspondence analysis lies in displaying such patterns for two-way tables of any size. If there is an association between the row and column ariables--that is, if the chi-square alue is significant--correspondence analysis may help reeal the nature of the relationship. Multiple Correspondence Analysis Multiple Correspondence Analysis tries to produce a solution in which objects within the same category are plotted close together and objects in different categories are plotted far apart. Each object is as close as possible to the category points of categories that apply to the object. In this way, the categories diide the objects into homogeneous subgroups. Variables are considered homogeneous when they classify objects in the same categories into the same subgroups. For a one-dimensional solution, multiple correspondence analysis assigns optimal scale alues (category quantifications) to each category of each ariable in such a way that oerall, on aerage, the categories hae maximum spread. For a two-dimensional solution, multiple correspondence analysis finds a second set of quantifications of the categories of each ariable unrelated to the first set, attempting again to maximize spread, and so on. Because categories of a ariable receie as many scorings as there are dimensions, the ariables in the analysis are assumed to be multiple nominal in optimal scaling leel. Multiple correspondence analysis also assigns scores to the objects in the analysis in such a way that the category quantifications are the aerages, or centroids, of the object scores of objects in that category. Relation to other Categories procedures. Multiple correspondence analysis is also known as homogeneity analysis or dual scaling. It gies comparable, but not identical, results to correspondence analysis when there are only two ariables. Correspondence analysis produces unique output summarizing the fit and quality of representation of the solution, including stability information. Thus, correspondence analysis is usually preferable to multiple correspondence analysis in the two-ariable case. Another difference between the two procedures is that the input to multiple correspondence analysis is a data matrix, where the rows are objects and the columns are ariables, while the input to correspondence analysis can be the same data matrix, a general proximity matrix, or a joint contingency 8 IBM SPSS Categories 24

13 table, which is an aggregated matrix in which both the rows and columns represent categories of ariables. Multiple correspondence analysis can also be thought of as principal components analysis of data scaled at the multiple nominal leel. Relation to standard techniques. Multiple correspondence analysis can be thought of as the analysis of a multiway contingency table. Multiway contingency tables can also be analyzed with the Crosstabs procedure, but Crosstabs gies separate summary statistics for each category of each control ariable. With multiple correspondence analysis, it is often possible to summarize the relationship between all of the ariables with a single two-dimensional plot. An adanced use of multiple correspondence analysis is to replace the original category alues with the optimal scale alues from the first dimension and perform a secondary multiariate analysis. Since multiple correspondence analysis replaces category labels with numerical scale alues, many different procedures that require numerical data can be applied after the multiple correspondence analysis. For example, the Factor Analysis procedure produces a first principal component that is equialent to the first dimension of multiple correspondence analysis. The component scores in the first dimension are equal to the object scores, and the squared component loadings are equal to the discrimination measures. The second multiple correspondence analysis dimension, howeer, is not equal to the second dimension of factor analysis. Multidimensional Scaling The use of Multidimensional Scaling is most appropriate when the goal of your analysis is to find the structure in a set of distance measures between a single set of objects or cases. This is accomplished by assigning obserations to specific locations in a conceptual low-dimensional space so that the distances between points in the space match the gien (dis)similarities as closely as possible. The result is a least-squares representation of the objects in that low-dimensional space, which, in many cases, will help you further understand your data. Relation to other Categories procedures. When you hae multiariate data from which you create distances and which you then analyze with multidimensional scaling, the results are similar to analyzing the data using categorical principal components analysis with object principal normalization. This kind of PCA is also known as principal coordinates analysis. Relation to standard techniques. The Categories Multidimensional Scaling procedure (PROXSCAL) offers seeral improements upon the scaling procedure aailable in the Statistics Base option (ALSCAL). PROXSCAL offers an accelerated algorithm for certain models and allows you to put restrictions on the common space. Moreoer, PROXSCAL attempts to minimize normalized raw stress rather than S-stress (also referred to as strain). The normalized raw stress is generally preferred because it is a measure based on the distances, while the S-stress is based on the squared distances. Multidimensional Unfolding The use of Multidimensional Unfolding is most appropriate when the goal of your analysis is to find the structure in a set of distance measures between two sets of objects (referred to as the row and column objects). This is accomplished by assigning obserations to specific locations in a conceptual low-dimensional space so that the distances between points in the space match the gien (dis)similarities as closely as possible. The result is a least-squares representation of the row and column objects in that low-dimensional space, which, in many cases, will help you further understand your data. Relation to other Categories procedures. If your data consist of distances between a single set of objects (a square, symmetrical matrix), use Multidimensional Scaling. Relation to standard techniques. The Categories Multidimensional Unfolding procedure (PREFSCAL) offers seeral improements upon the unfolding functionality aailable in the Statistics Base option (through ALSCAL). PREFSCAL allows you to put restrictions on the common space; moreoer, PREFSCAL attempts to minimize a penalized stress measure that helps it to aoid degenerate solutions (to which older algorithms are prone). Chapter 1. Introduction to Optimal Scaling Procedures for Categorical Data 9

14 Aspect Ratio in Optimal Scaling Charts Aspect ratio in optimal scaling plots is isotropic. In a two-dimensional plot, the distance representing one unit in dimension 1 is equal to the distance representing one unit in dimension 2. If you change the range of a dimension in a two-dimensional plot, the system changes the size of the other dimension to keep the physical distances equal. Isotropic aspect ratio cannot be oerridden for the optimal scaling procedures. 10 IBM SPSS Categories 24

15 Chapter 2. Categorical Regression (CATREG) Categorical regression quantifies categorical data by assigning numerical alues to the categories, resulting in an optimal linear regression equation for the transformed ariables. Categorical regression is also known by the acronym CATREG, for categorical regression. Standard linear regression analysis inoles minimizing the sum of squared differences between a response (dependent) ariable and a weighted combination of predictor (independent) ariables. Variables are typically quantitatie, with (nominal) categorical data recoded to binary or contrast ariables. As a result, categorical ariables sere to separate groups of cases, and the technique estimates separate sets of parameters for each group. The estimated coefficients reflect how changes in the predictors affect the response. Prediction of the response is possible for any combination of predictor alues. An alternatie approach inoles regressing the response on the categorical predictor alues themseles. Consequently, one coefficient is estimated for each ariable. Howeer, for categorical ariables, the category alues are arbitrary. Coding the categories in different ways yield different coefficients, making comparisons across analyses of the same ariables difficult. CATREG extends the standard approach by simultaneously scaling nominal, ordinal, and numerical ariables. The procedure quantifies categorical ariables so that the quantifications reflect characteristics of the original categories. The procedure treats quantified categorical ariables in the same way as numerical ariables. Using nonlinear transformations allow ariables to be analyzed at a ariety of leels to find the best-fitting model. Example. Categorical regression could be used to describe how job satisfaction depends on job category, geographic region, and amount of trael. You might find that high leels of satisfaction correspond to managers and low trael. The resulting regression equation could be used to predict job satisfaction for any combination of the three independent ariables. Statistics and plots. Frequencies, regression coefficients, ANOVA table, iteration history, category quantifications, correlations between untransformed predictors, correlations between transformed predictors, residual plots, and transformation plots. Categorical Regression Data Considerations Data. CATREG operates on category indicator ariables. The category indicators should be positie integers. You can use the Discretization dialog box to conert fractional-alue ariables and string ariables into positie integers. Assumptions. Only one response ariable is allowed, but the maximum number of predictor ariables is 200. The data must contain at least three alid cases, and the number of alid cases must exceed the number of predictor ariables plus one. Related procedures. CATREG is equialent to categorical canonical correlation analysis with optimal scaling (OVERALS) with two sets, one of which contains only one ariable. Scaling all ariables at the numerical leel corresponds to standard multiple regression analysis. To Obtain a Categorical Regression 1. From the menus choose: Analyze > Regression > Optimal Scaling (CATREG) Select the dependent ariable and independent ariable(s). Copyright IBM Corporation 1989,

16 3. Click OK. Optionally, change the scaling leel for each ariable. Define Scale in Categorical Regression You can set the optimal scaling leel for the dependent and independent ariables. By default, they are scaled as second-degree monotonic splines (ordinal) with two interior knots. Additionally, you can set the weight for analysis ariables. Optimal Scaling Leel. You can also select the scaling leel for quantifying each ariable. Spline Ordinal. The order of the categories of the obsered ariable is presered in the optimally scaled ariable. Category points will be on a straight line (ector) through the origin. The resulting transformation is a smooth monotonic piecewise polynomial of the chosen degree. The pieces are specified by the user-specified number and procedure-determined placement of the interior knots. Spline Nominal. The only information in the obsered ariable that is presered in the optimally scaled ariable is the grouping of objects in categories. The order of the categories of the obsered ariable is not presered. Category points will be on a straight line (ector) through the origin. The resulting transformation is a smooth, possibly nonmonotonic, piecewise polynomial of the chosen degree. The pieces are specified by the user-specified number and procedure-determined placement of the interior knots. Ordinal. The order of the categories of the obsered ariable is presered in the optimally scaled ariable. Category points will be on a straight line (ector) through the origin. The resulting transformation fits better than the spline ordinal transformation but is less smooth. Nominal. The only information in the obsered ariable that is presered in the optimally scaled ariable is the grouping of objects in categories. The order of the categories of the obsered ariable is not presered. Category points will be on a straight line (ector) through the origin. The resulting transformation fits better than the spline nominal transformation but is less smooth. Numeric. Categories are treated as ordered and equally spaced (interal leel). The order of the categories and the equal distances between category numbers of the obsered ariable are presered in the optimally scaled ariable. Category points will be on a straight line (ector) through the origin. When all ariables are at the numeric leel, the analysis is analogous to standard principal components analysis. Categorical Regression Discretization The Discretization dialog box allows you to select a method of recoding your ariables. Fractional-alue ariables are grouped into seen categories (or into the number of distinct alues of the ariable if this number is less than seen) with an approximately normal distribution unless otherwise specified. String ariables are always conerted into positie integers by assigning category indicators according to ascending alphanumeric order. Discretization for string ariables applies to these integers. Other ariables are left alone by default. The discretized ariables are then used in the analysis. Method. Choose between grouping, ranking, and multiplying. Grouping. Recode into a specified number of categories or recode by interal. Ranking. The ariable is discretized by ranking the cases. Multiplying. The current alues of the ariable are standardized, multiplied by 10, rounded, and hae a constant added so that the lowest discretized alue is 1. Grouping. The following options are aailable when discretizing ariables by grouping: Number of categories. Specify a number of categories and whether the alues of the ariable should follow an approximately normal or uniform distribution across those categories. 12 IBM SPSS Categories 24

17 Equal interals. Variables are recoded into categories defined by these equally sized interals. You must specify the length of the interals. Categorical Regression Missing Values The Missing Values dialog box allows you to choose the strategy for handling missing alues in analysis ariables and supplementary ariables. Strategy. Choose to exclude objects with missing alues (listwise deletion) or impute missing alues (actie treatment). Exclude objects with missing alues on this ariable. Objects with missing alues on the selected ariable are excluded from the analysis. This strategy is not aailable for supplementary ariables. Impute missing alues. Objects with missing alues on the selected ariable hae those alues imputed. You can choose the method of imputation. Select Mode to replace missing alues with the most frequent category. When there are multiple modes, the one with the smallest category indicator is used. Select Extra category to replace missing alues with the same quantification of an extra category. This implies that objects with a missing alue on this ariable are considered to belong to the same (extra) category. Categorical Regression Options The Options dialog box allows you to select the initial configuration style, specify iteration and conergence criteria, select supplementary objects, and set the labeling of plots. Supplementary Objects. This allows you to specify the objects that you want to treat as supplementary. Simply type the number of a supplementary object (or specify a range of cases) and click Add. You cannot weight supplementary objects (specified weights are ignored). Initial Configuration. If no ariables are treated as nominal, select the Numerical configuration. If at least one ariable is treated as nominal, select the Random configuration. Alternatiely, if at least one ariable has an ordinal or spline ordinal scaling leel, the usual model-fitting algorithm can result in a suboptimal solution. Choosing Multiple systematic starts with all possible sign patterns to test will always find the optimal solution, but the necessary processing time rapidly increases as the number of ordinal and spline ordinal ariables in the dataset increase. You can reduce the number of test patterns by specifying a percentage of loss of ariance threshold, where the higher the threshold, the more sign patterns will be excluded. With this option, obtaining the optimal solution is not garantueed, but the chance of obtaining a suboptimal solution is diminished. Also, if the optimal solution is not found, the chance that the suboptimal solution is ery different from the optimal solution is diminished. When multiple systematic starts are requested, the signs of the regression coefficients for each start are written to an external IBM SPSS Statistics data file or dataset in the current session. See the topic Categorical Regression Sae on page 15 for more information. The results of a preious run with multiple systematic starts allows you to Use fixed signs for the regression coefficients. The signs (indicated by 1 and 1) need to be in a row of the specified dataset or file. The integer-alued starting number is the case number of the row in this file that contains the signs to be used. Criteria. You can specify the maximum number of iterations that the regression may go through in its computations. You can also select a conergence criterion alue. The regression stops iterating if the difference in total fit between the last two iterations is less than the conergence alue or if the maximum number of iterations is reached. Label Plots By. Allows you to specify whether ariables and alue labels or ariable names and alues will be used in the plots. You can also specify a maximum length for labels. Chapter 2. Categorical Regression (CATREG) 13